How Green are Java Best Coding Practices? - GreenDays @ Rennes - 2014-07-01

How Green are Java Best Practices?Best Coding Practices and Software Eco-Design

Jérôme Rocheteau

Institut Catholique d’Arts et Métiers, Nantes, France

GreenDays@Rennes – Mardi 1er juillet 2014

How Green are Java Best Practices? GreenDays | 2014-07-01 1 / 33

mailto:[email protected]

Overview

1 Eco-Design and Best Practices

2 Formalizing Practices

3 Measuring Codes

4 Analyzing Measures, Codes, Practices

5 Eco-Design Indicators


Eco-Design and Best Practices


Context and Issues

Hypothesis and Objectives


3 Measuring Codes




Context and Issues

Context:

ICT accounted 2% of carbon emissions in 2007Energy efficiency relies on hardware but not softwareWorks on energy efficiency classes, energy-aware systems

Issues:

Help developers to build energy-efficient software1 Detect energy-consuming patterns in source code2 Replace these patterns by energy-saving ones

Qualify the energy impact for such best practices


Hypothesis and Objectives

HypothesisBest coding practices are eco-design rules

Objectives:

1 Formalizing Java best practices2 Measuring savings of memory and energy at runtime3 Evaluating confidence of such results


Formalizing Practices



Informal and Formal Examples

Time, Space, Energy Semantics

3 Measuring Codes





Best Coding Practice Example / Informal Rule:

String Literal InitializationInitialization of strings with the keyword ’new’ creates newobjects. This is costly in time and memory space. Initializationswith literal strings avoids this.

Three shape of Java best coding practices :1 Prefer this2 Avoid that3 Replace that by this

Can be extended to other primitive/wrapper typesCan be applied to other programming languages












Can be extended to other primitive/wrapper types

Can be applied to other programming languages















Listing 1: Prefer String Literal Initialization<rule id="prefer-string-literal-initialization">

<title>Prefer string literal initialization</title><description>

Primitive type objects should be initialized with primitive valuesand without the use of any constructors.

</description><check green="StringValue" gray="StringObject" />

</rule>



Listing 2: String Literal Initializationp u b l i c c l a s s S t r i n g V a l u e implements Code {

p r i v a t e S t r i n g [ ] a r r a y ;

p u b l i c vo id setUp ( ) {a r r a y = new S t r i n g [ 1 0 0 0 ] ;

}

p u b l i c vo id doRun ( ) throws Excep t i on {f o r ( i n t i = 0 ; i < 1000 ; i ++) {

a r r a y [ i ] = " abcde fg . . . " ;}

}

p u b l i c vo id tearDown ( ) {a r r a y = n u l l ;

}

}



Listing 3: String Object Initializationp u b l i c c l a s s S t r i n g O b j e c t implements Code {

p r i v a t e S t r i n g [ ] a r r a y ;

p u b l i c vo id setUp ( ) {a r r a y = new S t r i n g [ 1 0 0 0 ] ;

}

p u b l i c vo id doRun ( ) throws Excep t i on {f o r ( i n t i = 0 ; i < 1000 ; i ++) {

a r r a y [ i ] = new S t r i n g ( " abcde fg . . . " ) ;}

}

p u b l i c vo id tearDown ( ) {a r r a y = n u l l ;

}

}



Prefer LiteralInitialization

time = ft(td , tc)space = fs(sd , sc)energy = fe(ed , ec)



String ValueInitialization


String ObjectInitialization







time tcspace scenergy ec

time tdspace sdenergy ed







time tcspace scenergy ec

time tdspace sdenergy ed




Goal: statistical static analysis method and tools

original code




original code




original code refactored code




original code refactored code

possible savingsX joules / X %


Measuring Codes



3 Measuring Codes

Needs and Requirements

Measure Task and Process





Needs:

power-meter with digital outputsmemory monitor with digital outputsplatform for running Java codessystem for managing codes and measures

Requirements:

power-meter with fine-grain precisionJava micro-benchmarking to avoid JITsystem able to manage physical and logical sensors



Needs:

power-meter with digital outputs

memory monitor with digital outputsplatform for running Java codessystem for managing codes and measures

Requirements:

power-meter with fine-grain precision

Java micro-benchmarking to avoid JITsystem able to manage physical and logical sensors



Needs:

power-meter with digital outputsmemory monitor with digital outputs

platform for running Java codessystem for managing codes and measures

Requirements:

power-meter with fine-grain precision

Java micro-benchmarking to avoid JITsystem able to manage physical and logical sensors



Needs:

power-meter with digital outputsmemory monitor with digital outputsplatform for running Java codes

system for managing codes and measures

Requirements:

power-meter with fine-grain precisionJava micro-benchmarking to avoid JIT

system able to manage physical and logical sensors



Needs:

power-meter with digital outputsmemory monitor with digital outputsplatform for running Java codessystem for managing codes and measures

Requirements:

power-meter with fine-grain precisionJava micro-benchmarking to avoid JITsystem able to manage physical and logical sensors


Measure Task and ProcessMeasurement Protocol:

1 4 seconds idle2 10 seconds execution time3 3 seconds idle


Measure Task and ProcessSemantics based on quantitative metrics:

x-axis execution timey-axis instant power or instant memory space


Measure Task and ProcessPreliminary Computations:

total energy obtained by the trapezoidal ruleidle power = average of the first 4 second instant powersidle energy = idle power × protocol time



total energy obtained by the trapezoidal rule

idle power = average of the first 4 second instant powersidle energy = idle power × protocol time



total energy obtained by the trapezoidal ruleidle power = average of the first 4 second instant powers

idle energy = idle power × protocol time



total energy obtained by the trapezoidal ruleidle power = average of the first 4 second instant powersidle energy = idle power × protocol time


Measure Task and ProcessCode energy computation and normalization:

code energy = total energy - idle energy

normalized code energy = code energy / times of execution


Measure Task and ProcessCode energy computation and normalization:

code energy = total energy - idle energynormalized code energy = code energy / times of execution



Measure Process:

1 Retrieve an available Java code2 Iterate while this code isn’t mature:

Clean its measure setCompute its maturityPlan new measures if requiredPerform these measures

3 Send the report of this code



Measure Process:

1 Retrieve an available Java code

2 Iterate while this code isn’t mature:





Measure Process:






Measure Process:


Clean its measure set

Compute its maturityPlan new measures if requiredPerform these measures




Measure Process:


Clean its measure setCompute its maturity

Plan new measures if requiredPerform these measures




Measure Process:


Clean its measure setCompute its maturityPlan new measures if required

Perform these measures3 Send the report of this code



Measure Process:





Analyzing Measures, Codes, Practices



3 Measuring Codes


Clean and Canonical Measures

Code Maturity

Results

5 Eco-Design IndicatorsHow Green are Java Best Practices? GreenDays | 2014-07-01 17 / 33

Clean and Canonical Measures3 kinds of measure disturbances:

disturbances before the measure task underestimatedisturbances after the measure task overestimatedisturbances during the measure task overestimate



disturbances before the measure task underestimate

disturbances after the measure task overestimatedisturbances during the measure task overestimate



disturbances before the measure task underestimatedisturbances after the measure task overestimate

disturbances during the measure task overestimate



disturbances before the measure task underestimatedisturbances after the measure task overestimatedisturbances during the measure task overestimate



3 kinds of measure disturbances:

disturbances before the measure taskdisturbances after the measure taskdisturbances during the measure task

2 mere algorithms for cleaning measures:

bounds checking algorithm (over a single measure)split-and-merge algorithm (over a set of measures)






bounds checking algorithm (over a single measure)

split-and-merge algorithm (over a set of measures)






bounds checking algorithm (over a single measure)split-and-merge algorithm (over a set of measures)



Cleaning algorithm evaluation:

500 clean measures from 25 rules annotated by 3 experts

inter-agreement κ: 0.94 (almost perfect)baseline = quartile method

precision: 0.953recall: 0.911

split-and-merge algorithmprecision: 0.941recall: 1.0

remove all disturbed measures




500 clean measures from 25 rules annotated by 3 expertsinter-agreement κ: 0.94 (almost perfect)

baseline = quartile methodprecision: 0.953recall: 0.911






500 clean measures from 25 rules annotated by 3 expertsinter-agreement κ: 0.94 (almost perfect)baseline = quartile method
















split-and-merge algorithmprecision: 0.941recall: 1.0remove all disturbed measures


Clean and Canonical MeasuresCanonical measure:



time: t3space: s3energy: e3








time: tspace: senergy: e



















error margin: 2%


Code MaturityCode Maturity

How many clean measures required for reliable results?

0 20 40 60 80 100 120

5

10

15

Measure Number

Relativ

eStandard

Deviatio

n

a set of 25 measures minimuma standard deviation less than 10%




0 20 40 60 80 100 120

5

10

15

Measure Number

Relativ

eStandard

Deviatio

n





0 20 40 60 80 100 120

5

10

15

Measure Number

Relativ

eStandard

Deviatio

n

a set of 25 measures minimum

a standard deviation less than 10%




0 20 40 60 80 100 120

5

10

15

Measure Number

Relativ

eStandard

Deviatio

n



Results

Energy in nanojoules, time in nanosecondsMemory in kilobytesStandard Deviation on Energy

Replace Object Initialization by Literal Initialization

Rule Id Gree

nEn

ergy

Gray

Ener

gy

Gree

nT

ime

Gray

Tim

e

Gree

nM

emor

y

Gray

Mem

ory

Gree

nSt

dDev

Gray

StdD

ev

String 697 7885 842 7827 4136 36104 4.40% 8.14%Float 10311 10448 4736 4127 20032 20032 5.85% 6.58%Integer 685 9575 833 4591 4048 20032 5.21% 6.51%Boolean 683 6267 775 4741 4048 20032 4.77% 7.03%Char 695 33067 840 4595 4048 20032 5.04% 4.65%Double 10003 10210 5810 4270 28032 28032 5.58% 7.83%Long 669 8236 807 6066 4056 28032 2.89% 5.32%Short 680 7819 819 3846 4048 20031 4.87% 7.71%


Results

Replace Object Initialization by Literal Initialization

Rule Id Green Ene

rgy

GrayEne

rgy

Absolut

e Gain

Relativ

e Gain

String 697 7885 7188 91.16%Float 10311 10448 137 1.31%Integer 685 9575 8890 92.54%Boolean 683 6267 5584 89.10%Char 695 33067 32372 97.89%Double 10003 10210 207 2.02%Long 669 8236 7567 91.87%Short 680 7819 7139 91.30%


Results

Rules for loops

A Prefer integer loop countersB Avoid method loop conditionsC Prefer comparison-to-0 conditionsD Prefer first common condition

Id Gree

nEn

ergy

Gray

Ener

gy

Gree

nT

ime

Gray

Tim

e

Gree

nM

emor

y

Gray

Mem

ory

Gree

nSt

dDev

Gray

StdD

ev

A 82 98 4744 5931 16 16 8.66% 7.46%B 8376 8602 6181 6364 20056 20056 3.83% 6.14%C 89 284 4709 17367 16 16 5.09% 5.78%D 97 100 4934 5017 16 16 4.37% 4.60%


Results

Rules for loops

A Prefer integer loop countersB Avoid method loop conditionsC Prefer comparison-to-0 conditionsD Prefer first common condition

Id Green Ene

rgy

GrayEne

rgy

Absolut

e Gain

Relativ

e Gain

A 82 98 16 16.32%B 8376 8602 226 2.62%C 89 284 195 68.66%D 97 100 3 3.00%


ResultsSet String Builder or Buffer Size

A String Buffer/Builder capacity initializationB String Buffer/Builder capacity settingC String Buffer/Builder length setting

Id Gree

nEn

ergy

Gray

Ener

gy

Gree

nT

ime

Gray

Tim

e

Gree

nM

emor

y

Gray

Mem

ory

Gree

nSt

dDev

Gray

StdD

ev

A 593 1053 38085 59111 52056 73784 7.11% 8.40%B 598 1053 38495 59111 52056 73784 7.86% 8.40%C 1211 1053 59111 70765 73784 104064 8.40% 9.40%A’ 378 899 19278 41088 52056 73784 8.27% 7.18%B’ 429 899 41088 51214 73784 104064 7.18% 5.84%C’ 1043 899 20976 41088 52056 73784 6.01% 7.18%


Results

Set String Builder or Buffer Size

A String Buffer/Builder capacity initializationB String Buffer/Builder capacity settingC String Buffer/Builder length setting

Id Green Ene

rgy

GrayEne

rgy

Absolut

e Gain

Relativ

e Gain

A 593 1053 460 43.68%B 598 1053 455 43.20%C 1211 1053 -158 -15.00%A’ 378 899 521 57.95%B’ 429 899 470 52.28%C’ 1043 899 -144 -16.01%


Eco-Design Indicators



3 Measuring Codes



Summary

Future





HypothesisBest Coding Practices ≈ Eco-Design Rules

174 formalized and measured rules

55% rules huge savings (more than 10%)20% rules few savings (between 3% and 10%)15% rules no savings (between -3% and 3%)10% rules some losses (less than -3%)





55% rules huge savings (more than 10%)

20% rules few savings (between 3% and 10%)15% rules no savings (between -3% and 3%)10% rules some losses (less than -3%)





55% rules huge savings (more than 10%)20% rules few savings (between 3% and 10%)

15% rules no savings (between -3% and 3%)10% rules some losses (less than -3%)





55% rules huge savings (more than 10%)20% rules few savings (between 3% and 10%)15% rules no savings (between -3% and 3%)

10% rules some losses (less than -3%)





55% rules huge savings (more than 10%)20% rules few savings (between 3% and 10%)15% rules no savings (between -3% and 3%)10% rules some losses (less than -3%)



With The Way: 1 reliable measure system

hybrid physical and logical sensor managementcomplex client/server architecturefocused star schema database

robust automatic iterative processextensible programming API

reliable stable measure sets with few measures




hybrid physical and logical sensor management

complex client/server architecturefocused star schema database






hybrid physical and logical sensor managementcomplex client/server architecture

focused star schema databaserobust automatic iterative process

extensible programming APIreliable stable measure sets with few measures











robust automatic iterative process

extensible programming APIreliable stable measure sets with few measures















• Jérôme Rocheteau, Virginie Gaillard, et Lamya Belhaj.How Green are Java Best Coding Practices?Proceedings of the 3rd International Conference on SmartGrids and Green IT Systems,pages 235–246.Barcelona, Espagne, Avril 2014.


Future

1 Energy Efficiency Classes?

2 Eco-Design Indicators?

Absolute savingsRelative savings

3 Rule indicators in different environments4 Rule indicators in different programming languages5 Energy efficiency of different programming languages6 Accuracy of different physical and/or logical sensors7 Energy and memory footprints of logical sensors


Future

1 Energy Efficiency Classes?2 Eco-Design Indicators?




Future


Absolute savings

Relative savings3 Rule indicators in different environments4 Rule indicators in different programming languages5 Energy efficiency of different programming languages6 Accuracy of different physical and/or logical sensors7 Energy and memory footprints of logical sensors


Future





Future



3 Rule indicators in different environments

4 Rule indicators in different programming languages5 Energy efficiency of different programming languages6 Accuracy of different physical and/or logical sensors7 Energy and memory footprints of logical sensors


Future



3 Rule indicators in different environments4 Rule indicators in different programming languages

5 Energy efficiency of different programming languages6 Accuracy of different physical and/or logical sensors7 Energy and memory footprints of logical sensors


Future



3 Rule indicators in different environments4 Rule indicators in different programming languages5 Energy efficiency of different programming languages

6 Accuracy of different physical and/or logical sensors7 Energy and memory footprints of logical sensors


Future



3 Rule indicators in different environments4 Rule indicators in different programming languages5 Energy efficiency of different programming languages6 Accuracy of different physical and/or logical sensors

7 Energy and memory footprints of logical sensors


Future





Thank you


Data Model


Science

How Green are Java Best Coding Practices? - GreenDays @ Rennes - 2014-07-01